AI and the feedback problem no one talks about

There is a moment most people recognise, even if they rarely name it. You have looked at a piece of work and known, clearly, that it is not good enough. Maybe it misses the point entirely. Maybe it is confidently wrong. And then you have weighed up the cost of saying so, not because the feedback is unfair, but because of what it might do to the person on the receiving end. Whether they will take it personally. Whether it will damage the relationship. Whether it is worth the fallout.

That calculation happens constantly in workplaces. And it slows things down in ways that are hard to measure.

We already know how we talk to AI

Optimation recently put a question to its LinkedIn network: when AI gets something wrong, how do you respond? Of the people who voted, 75% said they tell it plainly and move on. Just 12.5% said they soften it the way they would with a person. Another 12.5% start over without any explanation at all.

That result is telling. Most of us, when no one is watching and no relationship is at stake, default to directness. We say what we mean. We do not manage the room. The question is what that reveals about how we give feedback the rest of the time, and what it costs us when we do not.

The 12.5% who start over without explanation are perhaps the most interesting group. They are not being kind, they are disengaging entirely. In a human context, that pattern looks like avoidance: sidestepping a difficult conversation by scrapping the work rather than addressing what is wrong with it. The fact that people do this with AI, where there is no relationship to protect, suggests the habit runs deep.

The case for radical honesty

Our Head of Digital Enablement, Scarlett, brought this up in a recent team discussion. A team member had described working with AI that produced project documentation but missed something important. They told it so directly, no softening, no diplomacy, no need to read the room. They got the honesty out without any of the usual human cost, then iterated from there.

What struck the group was not just the efficiency of that. It was what it removes from the equation. When a person gets significant feedback that their work has missed the mark, the emotional impact is real, and it often lingers. Disengagement, defensiveness, and in some cases churn, can all trace back to moments where feedback landed harder than intended. AI sidesteps that entirely. You can be as direct as the situation demands.

The before and after illustrates the point well. The first output was technically complete but contextually off, describing the right process for the wrong scenario. A softened response might have been: "this is a good start, could we maybe adjust the framing slightly?" What they said instead was closer to: "this is wrong, here is specifically why, rewrite it with these constraints." The revised output resolved the issue in a single pass. That difference of one round of iteration versus three or four compounds across a project.

Validation is still non-negotiable

That freedom to iterate bluntly comes with a responsibility the Optimation test team takes seriously. Neena D'Souza, one of Optimation's test leads, is clear that AI-generated outputs are always thoroughly validated before they are used. "AI can occasionally introduce invalid scenarios," she notes. The team uses AI to generate test scripts, documentation, and wiki pages, and the efficiency gains are real. But the human review step is not optional.

The feedback loop runs in both directions: the team holds AI to a high standard, and that is only possible because they have the domain knowledge to know what good looks like.

The reason that feedback was precise and effective is that the person giving it already knew what the documentation should say. The reason Neena's team catches invalid scenarios is that they understand the system being tested. Direct feedback without that underlying expertise is not radical honesty, it is just noise. The combination works precisely because of that. AI can absorb direct, unsoftened feedback and produce a better result quickly. But someone still needs to be sharp enough to spot what is wrong in the first place.

The flip side: when AI gives feedback to you

Most of the conversation about AI and feedback focuses on humans critiquing AI outputs. But the dynamic runs the other way too. AI tools are increasingly used to review code, flag issues in documents, suggest improvements to writing, and surface risks in plans. How teams respond to that feedback matters just as much.

The same patterns tend to emerge. Some people engage directly with what AI flags and act on it. Others soften the AI's suggestions before sharing them with the team, moderating the bluntness to protect feelings. Others dismiss AI feedback entirely because it came from a machine, even when the critique is valid.

The teams that get the most value from AI in a feedback role are the ones who treat it the same way they would want their own feedback treated: take what is useful, push back specifically on what is not, and do not let the source of the feedback determine whether it gets a fair hearing.

AI as a practice ground for harder conversations

There is also a less obvious benefit from this; teams that get comfortable giving direct, specific feedback to AI tools tend to get better at it with each other too.

Psychological safety, the sense that it is genuinely okay to speak up, is hard to build and easy to erode. But it is also a skill that can be practised. When people spend time in an environment where blunt, precise feedback is not just acceptable but actively effective, that muscle develops. The stakes with AI are low enough that the habit forms without the relational risk.

This is not a replacement for the harder work of building trust within a team. But it is a useful on-ramp. People who have spent time iterating directly and honestly with AI tools often find the language for feedback comes more naturally when they need it in a human context. They have already practised saying "this misses the point because…" rather than "this is really good, I just wonder if…"

What this means for how teams work

The deeper implication is about iteration speed and quality, not just comfort. When feedback is socially costly, people give less of it, and they give it later. Problems compound. Work goes further down the wrong path before anyone names the issue. AI does not eliminate the need for critical thinking, but it does remove one of the main reasons people hold back from applying it.

The teams at Optimation are not using AI to avoid hard conversations with each other. They are using it to have harder conversations with their work, earlier and more often, without the relational risk. That is a different kind of productivity gain, and it is one that does not show up in any time-tracking tool.

The risk to guard against is the inverse: teams that become so fluent in direct AI feedback that they forget the skill does not automatically transfer. Bluntness with a tool that cannot be hurt is easy. Directness with a colleague who has invested in their work is something else entirely. Both matter. The goal is not to outsource honest feedback to AI, but to use the practice to become better at giving it everywhere.

What does your team do when the feedback is hard to give?

If you are thinking about how AI fits into the way your teams work, not just the tools but the culture and the habits, we would like to talk. Reach out to our team here.

Contributors:

• Scarlett Maddock | Head of Digital Enablement
• Neena D'Souza | Test Lead